Dudley's martingale representation theorem states that if $W=\{W_t,\mathcal{F}_t;0\le t<+\infty\}$ is a standard one-dimensional Brownian motion, $0<T<+\infty$ and $\xi$ is $\mathcal{F}_T$-measurable, then there exists a progressively measurable process $Y=\{Y_t,\mathcal{F}_t;0\le t\le T\}$ satisfying $$\int_0^TY_t^2\mathrm{d}t<+\infty\text{ a.s.}$$ such that $$\xi=\int_0^TY_t\mathrm{d}W_t\text{ a.s.}$$ But this theorem does NOT have uniqueness. We need to find a progressively measureable process $Y$ such that $$0<\int_0^1Y_t^2\mathrm{d}t<+\infty\text{ a.s.,}\quad\text{but }\int_0^1Y_t\mathrm{d}W_t=0\text{ a.s.}$$

Could anyone give such an example?

  • $\begingroup$ Doesn't $Y = 1$ work? $\endgroup$ – Steve Huntsman Jan 12 '15 at 15:37
  • $\begingroup$ @SteveHuntsman: If $Y=1$ then $\int_0^1 Y_t\,dW_t = W_1$, not 0. $\endgroup$ – Nate Eldredge Jan 12 '15 at 17:55
  • $\begingroup$ Smack my head... $\endgroup$ – Steve Huntsman Jan 12 '15 at 23:34

You can find a counterexample in J. Michael Steele's Stochastic Calculus and Financial Applications on page 196. Here's an essentially equivalent construction.

Let $r(t)$ be any positive continuous function on $[0,1)$ with $\int_0^1 r(s)^2\,ds = +\infty$ (for example, $r(t) = 1/(1-t)$). Set $s(t) = \int_0^t r(s)^2\,ds$ so that $s$ is continuous on $[0,1)$, strictly increasing, and $s(1-) = +\infty$. If we let $Z_t = \int_0^t r(s)\,dW_s$ then $Z_t$ is a time-changed Brownian motion; specifically, $\{Z_{s^{-1}(u)}, 0 \le u < +\infty\}$ is a Brownian motion with respect to the filtration $\{\mathcal{F}_{s^{-1}(u)}, 0 \le u < +\infty\}$. (Observe that $Z_{s^{-1}(u)}$ is continuous with independent Gaussian increments having the correct variances.) Let $$\tau = \inf \left\{t \in \left[\frac{1}{2}, 1\right] : Z_t = 0\right\}.$$ Since Brownian motion is recurrent, almost surely $Z_{s^{-1}(u)}$ hits 0 for some $u \ge s^{-1}(1/2)$, thus $Z_t$ hits zero for some $t \ge 1/2$. So $\tau < 1$ almost surely. Finally set $$Y(t,\omega) = \begin{cases} r(t), & 0 \le t \le \tau(\omega) \\ 0, & \tau(\omega) < t \le 1. \end{cases}$$ Note $\int_0^1 Y_t^2\,dt = s(\tau)$. Since $\frac{1}{2} \le \tau < 1$ almost surely, by construction, we have $0 < s(1/2) < s(\tau) < \infty$ almost surely, hence $0 < \int_0^1 Y_t^2\,dt < \infty$ almost surely. And $\int_0^t Y_s\,dW_s = Z_{t \wedge \tau}$ so $\int_0^1 Y_s\,dW_s = Z_\tau = 0$.


To show non uniqueness it is sufficient to exhibit a non trivial process $\phi \in L^2_{loc}[0,T]$ such that $\int_{0}^{T} \phi \, dW =0$. This is because, in this case, the representation for $\xi$ is given by

$\xi = \int_{0}^{T} Y dW = \int_{0}^{T} Y dW + 0 = \int_{0}^{T} Y dW + \int_{0}^{T} \phi \, dW = \int_{0}^{T} (Y+\phi) \, dW $

and as $\phi$ is non trivial, we have $Y + \phi$ to be different to $Y$.

Our only restriction is that $\phi$ be (locally) square integrable - let us pick the example that Nate mentioned, as it is easy to understand.

Consider any real number $a$ in the open set $]0,T[$ and set

$\phi (t) = \mathbb{1}_{s \in [0,a[} + \frac{1}{(T-s)}\mathbb{1}_{s \in [a,\tau]}$

for $t \in [0,T]$, where $\tau$ is the stopping time

$\tau = \inf \lbrace t \geq a : \int_{a}^{t} \frac{1}{(T-z)} \, dW_z = -W_a \rbrace$.

Now, two important observations:

  • The event $\lbrace \tau < T \rbrace$ occurs $\mathbb{P}$-as, where $\mathbb{P}$ is the probability measure governed by the (standard) Weiner process $W$.
  • We have $\phi \in L^2_{loc}[0,T]$.

See that

$\int_{0}^{T} \phi \, dW = \int_{0}^{a} \phi \, dW + \int_{a}^{T} \phi \, dW = W_a - W_a = 0,$

therefore the element $Y+\phi$ is also a valid representation for $\xi$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.